Ahrefs Study on ChatGPT Citations: What Its Research Found

What are the main findings?

How was the study done?

Ahrefs content marketer Louise Linehan wrote the study with data scientist Xibeijia Guan, and Ryan Law reviewed it. The team analysed 1.4 million ChatGPT 5.2 prompts from desktop. The article dates them to February 2025, but ChatGPT 5.2 was released on December 11, 2025, so February 2026 is the likely date.

For each prompt, the data showed which URLs ChatGPT retrieved, which it cited, and the retrieval data returned with each result. Citing earlier research by Dan Petrovic, Ahrefs describes that data as a page title, a brief snippet, the URL and an ID number. Each URL was also tagged with an internal field, ref_type, showing the channel it came through. To estimate relevance, the team computed cosine similarity between page titles and both the prompt and ChatGPT's fan-out queries, the sub-questions it generates to find facts.

Ahrefs sells Brand Radar, the tool behind much of its AI data, and the article points readers to its reports. That does not make the findings wrong, but it is worth remembering when reading the recommendations.

What did it find about retrieval channels?

ChatGPT does not gather sources in one way. The ref_type field showed very different citation rates by channel:

• Search: 88.46% cited, across 25.6 million data points. This channel supplied 88% of all cited URLs.
• News: 12.01% cited, across 3.9 million data points.
• A dedicated community-discussion feed: 1.93% cited, despite more than 16 million data points. It made up 67.8% of all uncited URLs.
• Video: 0.51% cited.
• Academic: 0.40% cited.

Ahrefs's reading is that to be cited you need to be in the general search pool, which means your pages need to rank in the search results ChatGPT draws on. The discussion feed appears to be used as background: read for context, rarely credited.

Do titles and URLs affect citation?

In Ahrefs's data, yes. Cited pages had titles more similar to the user's prompt (0.602) than uncited pages (0.484). The best match between a cited page's title and ChatGPT's fan-out queries was higher still (0.656). Ahrefs concludes that matching the sub-questions ChatGPT asks behind the scenes matters more than matching the prompt someone typed.

URL slugs showed a smaller effect. Search results with natural-language slugs, such as /how-to-descale-a-coffee-machine, were cited 89.78% of the time, against 81.11% for results without them. Both groups are cited at high rates, so the slug is a modest signal, not a gate.

What did it find about freshness?

The picture is more mixed than "newer is better". In the search channel, the median cited page was about 500 days old, and some cited pages were more than 2,700 days old. Within a single prompt's retrieval set, it was often the newest pages that went uncited. Ahrefs's interpretation is that a new page is cited only if it matches the fan-out queries well; freshness alone does not win.

News was the exception. There, cited and uncited pages had nearly identical title relevance, and younger pages had a clear advantage. Freshness appears to act as a tie-breaker when relevance cannot separate the candidates.

What did it not find?

At first, uncited pages appeared to carry snippets and publication dates far more often than cited ones. Ahrefs traced most of that to the make-up of the data: the uncited pool was dominated by the discussion feed, whose items carry dates by default. Within the search channel alone, snippets were almost absent for both groups. Ahrefs's own conclusion is that it cannot draw strong conclusions about snippets or publication dates. It also warns that any cited-versus-uncited comparison that ignores where URLs came from risks mistaking quirks of the data for real patterns.

What do Ahrefs's other ChatGPT studies add?

• Most-cited pages (October 2025): of ChatGPT's 1,000 most-cited pages in September 2025, only 32.3% were pages a marketer could realistically pitch or influence. 28% had no organic search visibility, and those that ranked had a median Domain Rating of 90.
• Overlap with Google (September 2025): for short, head-term queries, 10% of ChatGPT's cited URLs were in Google's top 10, against 65% for Perplexity. At the domain level, overlap was 31.8%, so ChatGPT cites the same sites as Google about three times as often as the same pages.
• Top-10 overlap across assistants (August 2025): across 15,000 prompts, only 12% of links cited by ChatGPT, Gemini and Copilot ranked in Google's top 10 for the same prompt.
• Freshness (2025): across about 17 million citations, ChatGPT cited URLs 458 days newer than Google's organic results, the strongest freshness preference of the platforms tested.
• Schema (May 2026): pages that added JSON-LD showed no measurable ChatGPT citation gain against matched controls.

The 90% claim about lower-ranked pages is examined separately in whether 90% of ChatGPT citations come from pages ranked 21 or lower.

How should you read these findings?

• They are correlations. Cited pages had more relevant titles; that does not prove rewriting titles causes citations.
• Similarity scores are estimates. Ahrefs used open-source embedding models to approximate how ChatGPT may judge relevance.
• ChatGPT changes. The study covers ChatGPT 5.2 on desktop, and later models may behave differently. Answers also differ between the app and the API, as covered in why ChatGPT API answers differ from the web app.
• "Rank to be cited" needs care. The search channel draws on search indexes, but Ahrefs's overlap studies show ChatGPT's picks rarely match Google's top 10. Being indexed and relevant matters more than a particular Google position.

We have not reproduced the study, and the figures above are Ahrefs's own. Its full write-up is on Ahrefs's blog.

What should you do with it?

• Write titles that answer specific sub-questions, not just the head topic. Look at the fan-out queries tracking tools show for your key prompts.
• Use descriptive slugs on new pages. Do not change existing URLs just for this, since redirects carry their own risks.
• Make sure your pages are indexed in the major search engines, including Bing.
• Improve established pages that already match the questions, rather than publishing thin new ones.
• For news and time-sensitive topics, publish quickly and keep dates accurate.

How to find which pages ChatGPT prefers over yours is covered in how to reverse-engineer competitor pages cited by ChatGPT, and how many sources ChatGPT usually cites is in whether ChatGPT always provides sources.

What might this look like in practice?

Consider an illustrative home-appliance retailer, Plovby, whose guide titled "Our Complete Coffee Machine Guide" was retrieved for descaling questions but rarely cited. Its tracking tool showed ChatGPT's fan-out queries included "how often to descale a coffee machine" and "vinegar vs descaler for coffee machines". The team split the guide into two focused pages with titles matching those questions, at /how-often-to-descale-a-coffee-machine and /vinegar-vs-descaler. They left the original URL in place as a hub. Over the next two months, ChatGPT cited the new pages for several descaling prompts where it had previously cited a competitor.

What do people ask about the Ahrefs ChatGPT study?

How many prompts did Ahrefs analyse?

1.4 million ChatGPT 5.2 prompts, published in April 2026.

What share of retrieved pages does ChatGPT cite?

About half, according to Ahrefs, though some uncited pages were probably never opened.

Does ChatGPT prefer new content?

Across all citations it skews fresher than Google, but within one answer, older, well-matched pages were often cited over the newest ones.

Do URL slugs matter for ChatGPT citations?

Natural-language slugs were cited 89.78% of the time, against 81.11% for others. It is a modest correlation.